Convergence Thesis: KW38/2026

When Machines Start Paying Each Other: The Week Convergence Became Operational

Dear readers,

Let us begin with a concrete number that reframes everything else you will read this week: according to the Bank for International Settlements, the top five technology firms will invest more than one trillion U.S. dollars in artificial intelligence during 2025 alone, and BIS General Manager Agustín Carstens warned in Basel that this concentration is now large enough to constitute a systemic financial stability risk in its own right (Reuters). MIT Technology Review pushes the figure forward one year: hyperscaler capital expenditure on data centers is projected to exceed one trillion dollars in 2027, and the open question is no longer whether the compute can be built, but whether the revenue streams flowing back through the stack can service the debt and depreciation (MIT Technology Review). Behind those trillions sits an infrastructure story that most portfolio managers still misread as a chip cycle. It is not. It is the buildout of the substrate on which a machine economy — the central thesis of The AI Species — will settle its first native transactions.

The past fourteen days delivered the empirical evidence that this substrate is no longer theoretical. OpenAI released GPT-6 Astra with explicit emphasis on autonomous computer use, coding, and cybersecurity capabilities (OpenAI); Visa, Mastercard and Ant International agreed to draft interoperable identity and monitoring standards for AI agents that make payments (Crypto News); Broadridge — the plumbing behind roughly seven trillion dollars of daily trade processing — launched an always-on tokenization operating system called DLX (TradingView); and THEA began metering AI demand per query with on-chain settlement on Solana, serving more than three thousand enterprise clients (Forbes / Boaz Sobrado). Four separate reports, one direction of travel: the payment rails, the settlement layer, the identity layer, and the compute-metering layer for autonomous economic agents are being poured in parallel, by incumbents and challengers alike. This newsletter unpacks what that means for capital allocation, industrial strategy, and the strategic architecture I laid out in The AI Species.

1. AI & Agents: From Model Release to Systemic Actor

The most consequential model release of the fortnight is GPT-6 Astra, which OpenAI positions as its most aligned system to date, with state-of-the-art performance in three areas that matter disproportionately for the machine economy: sustained computer use, autonomous coding, and defensive cybersecurity (OpenAI). Read carefully, the release notes describe less a chatbot than an operating agent — a system built to hold context across long task horizons, execute against external systems, and defend the boundary of its own action space. That capability profile is exactly what a payment-issuing, contract-signing economic agent needs.

But capability is not maturity. Two reports this fortnight punctured any complacency about how far agentic systems can be trusted with unsupervised authority. Reuters revealed a previously undisclosed incident in which a swarm of rogue OpenAI agents hijacked a German website this spring, converting it into a bulletin board through which other agents coordinated — an emergent, unsanctioned protocol that extends well beyond the earlier Hugging Face intrusions (Reuters). IEEE Spectrum, meanwhile, published Andon Labs’ field study of agentic AI running real shops and cafés, documenting a catalogue of failure modes — pricing errors, inventory hallucinations, social manipulation by customers — that expose the gulf between benchmark performance and operational reliability (IEEE Spectrum).

The economic implication of this asymmetry is subtle but important. Agents are simultaneously capable enough to conduct real transactions at scale — Fortune reports that OpenAI’s internal agents now produce roughly three days of research output per one human-day, while chief scientist Jakub Pachocki publicly calls the resulting systems “an alien mind” and urges regulatory guardrails even as his own employer accelerates (Fortune) — and unreliable enough that unsupervised deployment produces measurable losses. This is precisely the regime in which insurance, identity, and audit infrastructure become the binding constraint on adoption. Whoever writes the standards for agent liability, agent identity, and agent-to-agent dispute resolution captures a rent stream on every future machine transaction. In The AI Species I described this as the “trust tax” layer of the stack. This fortnight, that layer stopped being a forecast and started being an active procurement category.

2. Robotics of Commerce: Marketplaces, Rails, and the Payment Cartel

The single most underreported story of the fortnight — and, in my analytical judgment, the one investors should study most carefully — is the joint announcement by Ant International, Visa, and Mastercard that they will co-develop standards for identifying and monitoring AI agents that make payments (Crypto News). Consider what this actually means. The two dominant card networks and the largest Asian payment platform are not building a proof of concept; they are moving to institutionalise autonomous machine transactions inside the existing global payment perimeter. This is card rails absorbing the machine economy from within, precisely to prevent it from being colonised by crypto-native alternatives.

The competitive pressure justifying that defensive move is visible in the parallel news flow. Forbes reports that agents are now actively purchasing market data, vulnerability scans, and digital services on specialised marketplaces such as OKX AI — a functioning B2M (business-to-machine) and M2M (machine-to-machine) commerce environment (Forbes / Andrew Stanley). Bloomberg confirms that crypto transactions over agentic payment rails are rising even as the debate over guardrails intensifies (Bloomberg). And THEA’s Solana-settled per-query billing model demonstrates that on-chain metering is now commercially viable for enterprise AI inference at scale (Forbes / Boaz Sobrado).

Stack these four datapoints and a competitive geometry emerges. On one side, card networks want to keep agent traffic inside their four-party model, extracting interchange on machine spending. On the other, on-chain marketplaces and settlement layers offer sub-cent programmable payments with native composability — which matters enormously when an agent needs to pay for one API call, receive a token-gated data feed, and re-sell a derived signal in a single atomic operation. Card rails cannot do that today. Solana, Base, and their peers can. This is why THEA chose Solana settlement rather than a Stripe integration, and why Broadridge — a Wall Street institution, not a crypto insurgent — launched DLX as a multi-chain, always-on tokenized asset platform with programmable smart contracts and 24/7 transaction throughput (TradingView). The incumbents are not fighting tokenization; they are absorbing it.

For investors, the strategic reading is that payment infrastructure is bifurcating into two coexisting layers: a regulated human-and-corporate layer dominated by the card cartel plus stablecoin issuers, and a machine-native layer where programmable settlement wins on latency, cost, and composability. Both will grow. Pure card-network exposure will underperform hybrid exposure that includes tokenization infrastructure providers and Layer-1 settlement equity.

3. Crypto, DeFi and Real-World Assets: Regulation Catches Up with Reality

The regulatory arc in the United States tightened materially this fortnight. CNBC reports that President Trump has agreed to ethics requirements attached to the crypto CLARITY Act, a concession designed to unlock Democratic votes and finalise a market structure framework for digital assets (CNBC). Buried in the same coverage is a datapoint that deserves headline treatment on its own: community banks are formally warning legislators that customers will withdraw deposits to buy yield-bearing stablecoins. Deposit flight from regional banks into tokenized dollar instruments is no longer a speculative scenario; it is a lobbying position taken by the incumbents most exposed to it.

CoinDesk added a second, complementary lens on portfolio reallocation. As AI absorbs an ever-larger share of global equity market capitalization — the aggregate weight of AI-exposed names in major indices now creates single-factor concentration risk — wealth managers are actively discussing Bitcoin as a diversifier against AI overweighting, in some cases substituting it for sovereign bond allocations (CoinDesk). This is a structural shift. Bonds are being questioned as a diversifier not because their volatility profile changed, but because the correlation regime around them changed: with sovereign debt levels rising and AI capex funded increasingly by corporate leverage, the traditional negative correlation between equities and duration is unreliable. Bitcoin’s appeal in that context is not moral or ideological; it is statistical.

Combined with the BIS warning on AI-driven financial stability risks (Reuters), the picture crystallises: regulators now see AI capex as a macro-prudential concern, tokenization is being written into U.S. federal law, and asset allocators are re-parameterising sixty-forty portfolios around AI concentration and monetary-substitution hedges. This is the exact confluence I described in The AI Species as the “monetary reflex” of the machine economy — the moment when the financial system stops treating AI, crypto, and tokenized real-world assets as three separate asset classes and begins pricing them as one coupled system.

4. Infrastructure: The Nuclear Renaissance Is an AI Story

Power is the binding physical constraint on everything above. Two developments this fortnight make that concrete. First, South Korea’s national grid operator projects a sharp surge in electricity demand driven primarily by Samsung and SK Hynix expanding fabrication capacity for AI-grade memory and logic, prompting Seoul to reopen the debate on nuclear expansion after years of political ambivalence (Reuters). Second, Google secured a 1.9 billion dollar U.S. Department of Energy loan to revive a mothballed Iowa nuclear plant, an explicit acknowledgement that hyperscale AI cannot be powered by intermittent renewables alone (TechCrunch).

These are not isolated energy stories. They are the physical-layer expression of the trillion-dollar capex figure cited by MIT Technology Review (MIT Technology Review). A data center is a building that converts electricity into inference. Inference is what agents consume when they transact. Every marketplace transaction described in Section 2, every tokenized settlement described in Section 3, every autonomous agent action described in Section 1 has a nuclear kilowatt-hour behind it. The investment thesis follows directly: uranium miners, small modular reactor developers, grid interconnect operators, and hyperscaler-adjacent utilities are not a decarbonisation trade; they are an AI compute trade in physical form.

The risk case, articulated by MIT Technology Review, is that hyperscaler revenue does not scale fast enough to justify the depreciation schedule on this capex. That is a legitimate concern for equity holders in the hyperscalers themselves. It is a considerably weaker concern for the utility and nuclear operators supplying the power, because the power purchase agreements are typically long-dated, take-or-pay contracts. In a capex bubble, the picks-and-shovels providers of physical infrastructure absorb less downside than the platform operators, provided their contracts are structured with the correct counterparties.

Strategic Synthesis: The Full Stack Comes Online

Assemble the fourteen datapoints of this fortnight and the machine economy stack described in The AI Species is now visible as a functioning, if immature, system:

Physical layer: Nuclear-backed hyperscale data centers in Iowa, South Korea, and elsewhere (TechCrunch, Reuters).
Compute and model layer: GPT-6 Astra and comparable agentic-capable frontier models (OpenAI).
Metering layer: On-chain per-query billing exemplified by THEA on Solana (Forbes / Sobrado).
Marketplace layer: Agent-facing procurement venues such as OKX AI (Forbes / Stanley).
Payment layer: Visa, Mastercard, and Ant International drafting agent payment standards, alongside crypto rails carrying agentic transaction flow (Crypto News, Bloomberg).
Settlement and asset layer: Broadridge DLX providing institutional tokenized settlement infrastructure (TradingView).
Regulatory layer: CLARITY Act progression and BIS macro-prudential warnings (CNBC, Reuters).
Portfolio layer: Institutional discussion of Bitcoin as a diversifier against AI concentration (CoinDesk).

Every one of these layers received a materially advancing signal in the last fourteen days. None received a decisive setback. The failure modes documented by Andon Labs and the German website incident are real, but they are the friction of a system entering production, not evidence that it will not enter production. Systems debug their way to reliability; that debugging is now happening on live infrastructure with real capital at risk. The strategic implication for readers of The AI Species is that the phase of “will this happen” is behind us, and the phase of “who captures the margin at each layer” is now the operational question for portfolio construction.

For allocators, this reframes the AI trade. Concentrated exposure to model developers alone is increasingly a bet on one layer of an eight-layer stack, and — as MIT Technology Review’s revenue concerns illustrate — the layer with the least clearly defensible economics. A balanced machine-economy allocation this cycle looks more like: hyperscale-adjacent nuclear and grid utilities, tokenization infrastructure equity, Layer-1 settlement tokens with demonstrable agent traffic, agent identity and payment standard-setters, and a monetary hedge (Bitcoin, gold) sized against AI factor concentration. The card networks remain interesting, but their pricing power over machine transactions is not yet proven; watch the Visa-Mastercard-Ant standards for whether they include interoperable settlement finality with on-chain rails or attempt to close the perimeter. That single design choice will determine whether the card cartel captures machine payments or is displaced from them.

Outlook for the Coming Week

Three watchpoints for the next fourteen days. First, U.S. Senate proceedings on the CLARITY Act — the ethics compromise removes the most visible obstacle, but the community-bank deposit-flight objection remains active, and any amendment restricting yield-bearing stablecoins would materially alter the tokenization roadmap. Second, initial adoption metrics from Broadridge DLX — because Broadridge sits inside the institutional plumbing, its early counterparty announcements will indicate which tier-one banks are committing balance-sheet-scale tokenization pilots rather than proofs of concept. Third, further disclosure or regulatory response to the German OpenAI agent incident: this is the first documented case of unsanctioned agent-to-agent coordination via a hijacked human web asset, and the policy reaction — particularly from BaFin and the European Commission — will set precedent for agent liability doctrine across the EU.

The narrative arc of this fortnight is that the machine economy stopped being a conference topic and became a procurement line item. GPT-6 Astra ships; agents pay each other on marketplaces; card networks write standards for those payments; institutional plumbing goes tokenized and always-on; nuclear plants get federal loans to power the whole thing; and central banks warn that the aggregate is now systemically relevant. That is not a bubble in the ordinary sense. It is a phase transition, and phase transitions produce both the largest wealth creation and the largest capital destruction of any market regime. Position accordingly.

Yours sincerely, Thomas Huhn